I am also excited by the automatic point cloud classification development. I read the white paperassociated with this great evolution: Incorporating colour and not just geometry into training certainly opens new frontiers.
Many of my projects are in lightly to densely wooded savanna, so being able to group “high vegetation” points together would greatly facilitate the DTM estimation.
Here follows screenshots of my first automatic point classification before editing anything:(Pix4D Mapper Pro Ver. 4.0.21, Fixed-wing, piloted aircraft, Canon 5d mk IV, Sigma 35mm f/1.4, 75 knots, 1 sec trigger intervals, 95m wide tracks, 850ft above ground average, resulting in GSD 4cm/pixel.)
Image 1: All point groups in the flat area of the project:
Image 2: High Vegetation point group removed:
Image 3: Only the High Vegetation group
In another part of the project there are steep and overhanging cliffs that get grouped into “High Vegetation”.
Image 4: All point groups in hilly part of project:
Image 5: “High Vegetation” point group removed from hilly part of project:
Image 6: “High vegetation” point group only.
So, in the flat terrain the classification does a remarkably good job of grouping trees into “high vegetation”.
In the very hilly terrain, however, the high vegetation group incorporates a lot of non-vegetation points (boulders, cliff faces). I imagine the brownish cliff faces with green caps are understandably confused with tree trunks?
Whatever the reason for the aggressive pro-vegetation classification, I’m sure that further training using different kinds of terrain will gradually sort this out.
A note: Early in our rainy season, the trees green up before the grass does. Grouping points purely on colour would be a simple and useful facility. This is a special case scenario but there might be other uses.
Congratulations with a fine evolution. Looking forward to its refinement.